Author: Kevin Heleodoro Date: February 24, 2024 Purpose: Notes on Project 3 for CS 5530
Options for thresholding values:
-
Find a value in the valley between two peaks of an image's histogram.
-
Pick a mid-range value of 128
-
Calculate the mean/median pixel value of the image.
-
Otsu's method - minimizes intra-class variance
-
Adaptive thresholding - calculates locally for different regions of the image.
-
https://docs.opencv.org/4.x/d7/d4d/tutorial_py_thresholding.html
-
https://docs.opencv.org/4.x/da/d97/tutorial_threshold_inRange.html
Compare the result if applying a gaussian blur to the image. How much of a benefit does it provide?
"Type: \n 0: Binary \n 1: Binary Inverted \n 2: Truncate \n 3: To Zero \n 4: To Zero Inverted";
type 1
value 70
type 1
value 70
type 1
value 152
type 1
value 70
Clean up the binary image:
- Use morphological filtering to clean up the images.
- noise reduction - morphological opening - white spots
- filling holes - morphological closing - black spots
I will use the filling holes approach since the thresholded images seem to be displaying a lot of black spots. What are the black spots?
- https://towardsdatascience.com/image-processing-class-egbe443-6-morphological-filter-e952c1ec886e
- https://docs.opencv.org/4.x/db/df6/tutorial_erosion_dilatation.html
- https://docs.opencv.org/4.x/d3/dbe/tutorial_opening_closing_hats.html
- https://medium.com/nattadet-c/morphological-filters-d99860d39b85
Kernel size of 5 on morph_close did not produce any changes to the black spots present in sunglasses_1 No changes when using the morph_open with kernel size 5.
Setting a kernel size of 25 filled in the gaps and overlapped into the shadows.
type 1
value 70
kernel size 25
type 1
value 70
kernel size 15
Differences between 25 and 15 seem negligible. However, dropping down to kernel size 10 allowed for the black spots to show again.
value 152
kernel size 15
The fork needs a lot more fine tuning. This is most likely due to the metallic aspects of it. There are many areas of light and dark in the same place.
value 70
kernel size 15
Segment into regions
-
Use the built-in
connectedComponentsWithStatsfunction to determine the regions within the image. -
https://docs.opencv.org/4.x/d2/dbd/tutorial_distance_transform.html
-
https://docs.opencv.org/3.4/d3/dc0/group__imgproc__shape.html#ga107a78bf7cd25dec05fb4dfc5c9e765f
value 70
kernel size 15
min area 250
value 126
kernel size 3
min area 250
value 91
kernel size 6
min area 250
Compute features for a region
- Calculate the least central moment and oriented bounding box
- percent filled
- bounding box
- https://docs.opencv.org/4.x/d0/d49/tutorial_moments.html